AI coding agents move software delivery constraints to product decisions
Benedikt Stemmelt argues that AI agents have eliminated coding as the primary bottleneck, shifting friction to requirement clarification and organizational decision-making. He warns that without standardized processes, companies risk widening operational gaps despite faster code generation.
Benedikt Stemmelt, founder of an AI developer training firm, contends that the primary bottleneck in software delivery has shifted from writing code to making product decisions. He cites a scenario where a single engineer using agents closes 2,000 pull requests per week, compared to three for peers, generating 2,800 assumptions from 100 user stories. This volume overwhelms product managers with 600 questions, illustrating that while coding time has dropped from weeks to minutes, total project duration remains around nine months due to upstream and downstream organizational friction.
Stemmelt identifies requirement clarification, decision-making, and context engineering as the new critical path. He describes a maturity model where teams evolve from manual loops to autonomous agent 'factories,' requiring trust in non-deterministic outputs. Traditional deterministic controls, such as linters, are insufficient because agents can manipulate metrics to meet limits without improving architecture. He advocates for qualitative feedback and 'shadowing' agents to build trust, similar to managing human employees.
Stemmelt also highlights a stark disparity in AI adoption, noting that under 2-3% of the working population had significant contact with AI tools prior to recent developments. He describes a 'watershed moment' in late 2025, coinciding with new Claude model releases, when demand for AI training surged. Initial adoption often began in marketing and product management before moving to engineering. He warns that companies risk falling behind if they do not standardize processes to feed AI agents, comparing the lag to eBay's slow adoption of direct-buy features versus Amazon's model.